From Code to Motion: Building an Autonomous Hat-Hunting Robot with Kubernetes & ML

What can a hat-hunting robot teach us about managing thousands of disconnected edge devices? Learn how cloud-native patterns extend all the way to the far edge.

#1about 3 minutes

Understanding the challenges of edge computing deployments

DevOps principles can be extended from the data center to manage workloads on disconnected or intermittently connected edge devices.

#2about 2 minutes

Introducing the robot's hardware and software stack

The robot is built on a Raspberry Pi running MicroShift, a lightweight Kubernetes distribution, and exposes a simple Flask REST API for motion control.

#3about 4 minutes

Designing the end-to-end system architecture

The system uses a central OpenShift cluster for development and model training, with Skupper for secure communication and ArgoCD for GitOps-based deployment to the robot.

#4about 7 minutes

Training an object detection model with OpenShift AI

A JupyterLab workbench is used to define and run an Elyra pipeline that trains a YOLOv5 model on the Open Images dataset to recognize fedora hats.

#5about 4 minutes

Deploying the trained model as an inference service

The trained ONNX model is deployed as a scalable and secure REST API endpoint using the model serving feature in OpenShift AI.

#6about 7 minutes

Developing the robot control application in a web IDE

A Python Flask application is developed using a web-based IDE (Eclipse Che) with a devfile to manage the workspace and connect to the inference service.

#7about 2 minutes

Live demonstration of the autonomous hat-hunting robot

The robot successfully uses its camera and the ML model to detect a red hat, calculate its position, and navigate towards it in real-time.

#8about 1 minute

Managing edge deployments with GitOps using ArgoCD

ArgoCD manages the robot as a remote Kubernetes cluster, enabling automated, Git-driven rollouts of new application and model versions to the edge device.